Anaphora and Semantic Interpretation:
A Reinterpretation of Reinhart's Approach*
Irene Heim
MIT
Syntactic structure constrains the possible "coreference" relations between the
noun phrases in a sentence. Since the early days of generative syntax, we have
learned a great deal about the nature of these structural constraints. Much of this
syntactic theorizing has been possible with only a vague and informal
understanding of what it really is in semantic terms that is being constrained
here. Everyone agrees that it is not coreference in a literal sense, i.e., sameness
of referents. But to specify in positive and semantically precise terms what it is
instead has turned out to be non-trivial. Those authors that have given serious
attention to the semantic import of syntactic constraints on so-called
"coreference" have arrived at rather different conclusions. 1
A particularly elegant and well-elaborated proposal in this regard is due
to Tanya Reinhart (1983a; 1983b). I ts central thesis is that only one type of
"coreference" relation is syntactically represented and directly constrained by
principles of grammar, and this is the well-understood relation of variable
binding in the sense of formal logic. Other semantic relations, in particular such
as may obtain between two referring NPs, are not even represented on any
syntactic level and can therefore not be directly licensed or ruled out by
structural conditions. They are regulated instead by an extragrammatical
principle which says, very roughly, that coreference is unavailable whenever the
same meaning can be conveyed by means of variable binding. We will get to
the details shortly.
The purpose of this article is two-fold: First, I will develop a
formulation of Reinhart's proposal that is in certain respects somewhat more
explicit than her own. Second, I will argue for a revision which, though it does
*This paper was written in 1992 and distributed in the working paper series of the
Seminar für Sprachwissenschaft of the University of Tübingen (SfS-Report-07-93). It is
reprinted here in its original form (except for typographical corrections and updated
publication dates in the bibliography).
Aside from its obvious debt to the work of Tanya Reinhart, the paper draws
heavily on insights that I obtained from the writings of James Higginbotham. Angelika
Kratzer's explication of Reinhart's theory in her 1990 lecture notes also influenced my
thoughts. James Higginbotham, Hans Kamp, Tony Kroch, Barbara Partee, Carl Pollard,
Arnim von Stechow, and Hiroaki Tada offered important comments on earlier
presentations of this material, and Tanya Reinhart has been extremely generous with oral
as well as written feedback for several years.
1The following works are helpful in clarifying the problems and contain a representative
sample of proposed solutions: Evans 1980; Higginbotham 1980 (note 1); Heim 1988
(ch. III, 2.3); Soames 1990; Fiengo & May 1994 (ch. 1).
The Interpretive Tract. MIT Working Papers in Linguistics 25, 205-246
U. Sauerland and O. Percus (eds.) © 1998 Irene Heim
Irene Heim
leave practically all of Reinhart's substantive insights intact, leads to a theory
which no longer fits the summary I just gave. In particular, the revised theory
implies that bound variable anaphora does not enjoy a special status, but
coreference anaphora is syntactically represented in exactly the same way.
The reader is warned, first, that it is not the purpose of this article to
survey or evaluate the competitors to Reinhart's approach which are already
found in the literature. Aside from some scattered allusions, they will be
disregarded. Of course, if there already is a successful alternative to Reinhart's
approach on the market, then this makes the present enterprise more or less
irrelevant. I do not believe that there is, but it would take a separate paper (or
several) to explain just why not. Second, I will also disregard the numerous
criticisms that other authors have already put forward against Reinhart.2 Some
of them, I believe, have been successfully countered or happen not to apply
against my particular formulation of her ideas. Others remain unrefuted, and
most of those will threaten the revised theory I endorse no less than Reinhart's
original version. For instance, I inherit what are likely to be the wrong
descriptive generalizations about Weak Crossover3 and about ellipsis 4. If I am
lucky, appropriate remedies for these and other shortcomings will not
undermine my main points, but for all I know they might. Apart from these
objections, which I am simply not competent to deal with, I neglect others for
mere reasons of space. In particular, I omit all discussion of Binding Condition
C, even though I defend claims that are not consistent with Reinhart's position
on this matter.5
1. Reinhart's theory
Reinhart's theory of the syntax and semantics of anaphoric relations is best
known from her book Anaphora and Semantic Interpretation (Reinhart 1983a).
My exposition mostly follows the more recent summary in Grodzinsky &
Reinhart 1993 (henceforth G&R). I take the liberty of making some small
technical changes to suit my personal tastes and habits, but these should not
distort any of the substantive ideas.
Let us begin with the central assumptions about the derivation and
well-formedness of S-structure (SS) . We have free, optional indexing . Any NP
may, but need not, be assigned an index (a numerical subscript), and different
2See especially Lasnik 1989 (ch. 9) and references cited there.
3For counterexamples and alternative proposals, see especially Higginbotham 1980 and
Stowell 1987.
4Reinhart and the present work basically follow Sag 1980, which is problematic in light
of a number of more recent studies (see especially Dalrymple, Schieber, Perreira 1991,
Kitagawa 1991, and Fiengo & May 1994.)
5I am persuaded that Condition C is required in the syntax in order to predict the
distribution of bound variable construals for epithets, as shown, e.g., by Haïk 1984
(204f.), Lasnik 1989 (ch. 9), and Higginbotham 1994. This point is independent of the
argument that I have with Reinhart in this article. Once Condition C is reintroduced
alongside A and B, most of what I say about Condition B below will probably carry over
mutatis mutandis to C.
206
Anaphora and Semantic Interpretation
NPs in a sentence may receive the same or different indices. For instance, we
can assign (1) any of the following indexings (among others 6), though (as we
will see below) some of these lead to derivations that are later filtered out as illformed or uninterpretable.
(1)
Every boy called his mother.
a.
every boy1 called his 1 mother
b.
every boy1 called his 2 mother
c.
every boy called his1 mother
d.
every boy1 called his mother
e.
every boy called his mother
One important filter consists of the Binding Theory conditions A and B, which
apply at SS:
(2)
Binding Conditions:
A.
B.
An anaphor is A-bound in its GC.
A pronominal is not A-bound in its GC.
(2) presupposes a lexical categorization into anaphors (in particular reflexive
pronouns) and pronominals (non-reflexive personal pronouns). It also
presupposes suitable characterizations of "Governing Category" (GC), for which
the reader is referred to the syntactic literature. Moreover, it relies on the
following definitions, which in turn appeal to standard definitions of "Aposition" and "c-command" (most of the time, it will not matter which particular
version).
(3)
a.
α binds β iff α c-commands and is coindexed with β.
b.
α A-binds β iff α binds β and α is in an A-position.
None of (1a-e) happen to be filtered out by the Binding Conditions, but the
reader is surely familiar with examples that would be.
From SS, a transformational derivation leads to Logical Form (LF) .
The main operation of interest in this derivation is so-called Quantifier Raising
(QR) . Contrary to what its name suggests, it applies optionally and freely to all
types of NPs. (But again, derivations in which QR has failed to apply will often
be ruled out by yielding uninterpretable outputs.) Specifically, QR is assumed
to apply in the following fashion: It replaces an indexed NP αi by a coindexed
trace, adjoins α (without the index!) to a dominating node, and prefixes the
sister constituent of α with a lambda operator indexed i. Schematically:
6For instance, I neglected the additional option of indexing the NP his mother.
207
Irene Heim
(4)
QR:
[S ... αi ... ] => [S α λi[S ... ti ... ] ]
For example, the result of applying QR to every boy 1 in (1a) is (1f).
(1)
f.
every boy λ1[t1 called his1 mother]
Our formulation implies that an NP needs to have an index in order to be able to
undergo QR; this means, e.g., that every boy cannot be QRed in (1c) or (1e).
This in itself presumably doesn't jeopardize the derivations, because quantifiers
in subject position are straightforwardly interpretable in situ, there being no
semantic type mismatch.
LFs are then submitted to the following definition of "variable" and the
associated filter.
(5)
a.
An index is a variable only if it is
(i) on a λ, or
(ii) on a trace and bound by a λ, or
(iii) on a pronominal or anaphor and A-bound.
b.
All indices must qualify as variables.
(5) cuts down on the number of possible derivations quite considerably. For one
thing, it implies that all (overt) NPs apart from anaphors and pronominals, in
particular all quantifiers and proper names, must wind up without an index.
This means they must either start out unindexed at SS, or else undergo QR and
thereby transfer their index to the λ. (Which makes QR effectively obligatory
for every boy1 in (1a,b,d).) From a semantic point of view, it makes sense not to
allow indices on quantifiers and names: on standard assumptions, the meanings
of such NPs are completely determined by the lexical entries for the words in
them and compositional rules; an index has no conceivable semantic
contribution to make and would thus have to be ignored anyway if it were
present.
(5) also implies that LFs cannot contain any free variables ("free" in the
sense of standard logic book definitions). For instance, the result of QRing only
every boy1 in (1b) is not a legitimate LF, because the index 2 is not sanctioned
by any clause of (5). (The only chance of rescuing this derivation would be by
QRing his2 as well.)
(1)
g.
* every boy λ1[t1 called his2 mother]
(5) moreover incorporates a version of the Weak Crossover prohibition, by
disallowing locally A-bar-bound pronouns. So the derivation (1a)/(1f) is wellformed, because the pronoun his1 winds up A-bound by the QR-trace t1, but its
counterpart with subject and object reversed would not be:
(6)
208
a.
SS:
his1 mother called every boy1
b.
LF:
* every boy λ1[his1 mother called t1]
Anaphora and Semantic Interpretation
The index on his in (6b) fails to count as a variable, because its only binder is
the λ rather than an A-position.
Finally, LFs are semantically interpreted . I can be informal about this
here, because how it works is mostly obvious. λ's, of course, are functional
abstractors. Variables in the sense defined in (5) are treated like variables in
logic (with occurrences of the same numerical index counting as occurrences of
the same variable, and occurrences of distinct numerical indices as occurrences
of distinct variables). A constituent that bears an index simply inherits the
interpretation of its index. For this reason, it is harmless and natural to refer to
the whole indexed NP as a "variable", though strictly speaking, the variable is
just the index.
Pronouns without indices are deictic and the utterance context has to
provide referents for them. Apart from sortal restrictions due to the pronoun's
gender, number, and person features, this reference assignment is a result of
interacting pragmatic factors, including salience and overall plausibility.
Moreover, it is constrained by the following principle, which constitutes the
most distinctive ingredient of Reinhart's approach.
(7)
Coreference Rule:
α cannot corefer with β if an indistinguishable interpretation can be
generated by (indexing and moving β and)7 replacing α with a variable
A-bound by the trace of β.
The next section is entirely devoted to illustrations of (7), which will also serve
to clarify some of the concepts it employs, notably "interpretation" and
"indistinguishability".
2. Reinhart's Coreference Rule applied to examples
This section parallels the discussion of the Coreference Rule in G&R, section
2.3. In particular, my example groups (ii) - (v) are all taken from their list, with
one systematic alteration: Since I am not dealing with Condition C effects at all
in this paper, I have replaced all apparent Condition C violations by similar
examples that look as though they violate Condition B.
2.1. Group (i): basic cases
Let's look at three primitive examples containing the proper name John and a
masculine singular pronoun.
(8)
John saw him .
(9)
John saw his mother.
(10)
His mother saw John.
7The parenthesized part of the instructions can be skipped if
β was already QRed in the
original structure.
209
Irene Heim
In each case, we are interested in the possible relations between the
interpretations of the name and the pronoun. This breaks down into two
questions: First, can the two stand in a variable binding relation? Second, can
they corefer? The predictions turn out to be the following: (8) allows neither
binding nor coreference; (9) allows binding but not coreference; and (10) allows
coreference but not binding. Here is how they are arrived at:
When we ask whether the him in (8) could be a variable bound by John,
we can't, of course, mean this quite literally; proper names are not variable
binders, after all. What we really mean is whether the pronoun could be a
variable bound by the λ that arose from QRing the name. So the question of
whether binding is possible in (8) turns on the well-formedness of the following
derivation.
(8)
a.
SS:
* John1 saw him1
b.
LF:
John λ1[t1 saw him1]
Though all indices in the LF (8b) qualify as variables and there is no obstacle to
interpretability, this derivation is already filtered out at SS by Binding Condition
B.
Could the two NPs corefer? For this we would need an LF as in (8c)
(trivially derived from an identical SS in which no NP was indexed) and an
utterance context that furnishes the reference assignments indicated by the
pointers underneath.
(8)
c.
LF:John saw him
↓
↓
j
j
I will make use of this notation to specify utterance contexts throughout the
paper: Each referring NP in the LF is connected by an arrow to its contextually
supplied referent. The lower-case letters stand for individuals out there in the
world, with each letter representing a unique individual and each individual
represented by a unique letter.
There would be nothing wrong with the interpretation indicated in (8c),
if it weren't for the Coreference Rule. This rule instructs us to look for an
alternative LF that results by certain specified alterations from that in (8c) and to
make sure that it wouldn't yield an indistinguishable interpretation. A potential
such alternative happens to be the LF we already saw in (8b), set in this context:
(8)
d.
LF:John λ1[t1 saw him1]
↓
j
This differs minimally from (8c) in just the way that (7) instructs us: Of our two
NPs in (8c) that were candidates for a coreferring pair, the first (John) has been
indexed and QRed and the second (his) has been replaced by a variable (his 1)
that is A-bound by the trace of the former. It is an interpretable LF,
notwithstanding the fact that it (because it would have to derive from a
210
Anaphora and Semantic Interpretation
Condition B violation at SS) is not part of a well-formed derivation for any
English sentence.8
Now we must determine whether there is any difference in the
interpretations of (8c) and (8d). Suppose for now this is to be done by
comparing the propositions expressed by each LF in its indicated context . It
turns out that these are the same: (8c) is true in any world where j saw j, and so
is (8d). So (8d) represents an indistinguishable interpretation from (8c), and
therefore the Coreference Rule says that (8c) is not a permissible interpretation
for (8). In short, the option of coreference in (8) is preempted by the existence
of (8d).
Let's turn to (9).
(9)
John saw his mother.
Here, binding is evidently predicted possible, the following derivation being
well-formed on both levels.
(9)
a.
SS:
John1 saw his1 mother
b.
LF:
John λ1[t1 saw his1 mother]
But coreference is not. A coreferential interpretation would look like this:
(9)
c.
LF:
John saw his mother
↓
↓
j
j
But under the Coreference Rule, this is preempted by (9d) (= (9b) plus a
context).
(9)
d.
LF:
John λ1[t1 saw his1 mother]
↓
j
Both of these express the proposition that j saw j's mother.
(As G&R acknowledge in footnote 13, it might be preferable to predict
this example to be ambiguous between a bound and a coreferential reading.
They suggest that this could be accomplished by a revision of (7) that confines
its application to those examples which involve prima facie violations of
Condition B, in a sense they make precise. I will adopt a similar proposal
below, but set the issue aside for the time being.)
8I have chosen to read (7) in such a way that the potential alternative structures to be
considered in applying this principle need not be part of complete grammatical
derivations. Alternatively, one might impose this further requirement, in which case the
him in (8c) would have to be replaced by a himself1 in (8d). Most renditions of Reinhart's
proposal seem to assume the latter. It doesn't seem to make any difference for the cases
considered here, but see footnote 12 of G&R (and apparently Reinhart (1991a), which I
haven't seen).
211
Irene Heim
In (10), we have the reverse prediction.
(10)
His mother saw John.
Binding is out, because (as we already saw with (6b)), the requisite structure
contains an index on his 1 that fails to qualify as a variable.
(10)
a.
* John λ1[his1 mother saw t1]
But coreference is, for this very reason, allowed. (10b) depicts the relevant
interpretation.
(10)
b.
LF:his mother saw John
↓
↓
j
j
This is not preempted by any other structure. Why not? Because the closest we
can come to constructing a potential competitor according to the specifications
of (7) is to index and QR John and coindex his with its trace, but then we have
precisely (10a), where his 1 is not a variable.
These three examples should have clarified some mechanical aspects of
the Coreference Rule. They also gave us the opportunity for a first stab at
elucidating the notion of indistinguishable interpretations, but we will soon see
that there is more to this notion than we have so far uncovered.
2.2. Group (ii): examples with only
(11) illustrates another type of example which has the superficial appearance of
a Condition B violation and which Reinhart cites in support of the Coreference
Rule. (I won't talk about analogous cases with other focussing particles such as
even.)
(11)
(Everybody hates Lucifer.) Only he himself pities him .
Why is (11a), with coreference, an available interpretation and not preempted by
the binding-alternative (11b)?
(11)
212
a.
only he himself λ1[t1 pities him]
↓
↓
l
l
b.
only he himself λ1[t1 pities him1]
↓
l
Anaphora and Semantic Interpretation
Well, that's obvious. We don't need to cast around for subtle differences here.
These two LFs have manifestly different truth conditions. (11a) says that
nobody besides Lucifer pities Lucifer , whereas (11b) claims that there are no
other self-pitiers.
Actually, a closer look at (11b) reveals a potential problem with this
application of (7). Notice that what corresponds to the α and β of (7) here are
the NPs him and he himself respectively. It is these two, after all, whose
coreference is to be licensed. When we construct the potential competing
structure according to the specifications of (7), we are supposed to coindex α
with the trace of β. But the trace t1 with which we coindexed him1 in (11b) is
not really the trace of he himself, but rather the trace of only he himself. That is
a different NP, and not a referring one, hence not a possible choice for β in the
first place.
In short, if this is indeed the way in which G & R intend the
Coreference Rule to apply to this example, then they must somehow be
assuming that for the purposes of (7), t 1 in (11b) counts as a trace of he himself.
That seems a little bit hokey, but there is something to be said for it. I will
return to the matter in section 5.3.3 below.
2.3. Group (iii): identity under debate
Coreference is systematically possible in Condition B configurations when we
are dealing either with explicit identity statements or with other utterances in
discourse contexts where the identity of the referents is unknown or at issue. A
representative example is the second to last sentence in (12).
(12)
A:
Is this speaker Zelda?
B:
How can you doubt it? She praises her to the sky. No competing
candidate would do that.
Nothing is wrong with this if the woman in question indeed is Zelda and the
pronouns thus corefer. Nothing is even wrong if the speaker knew this all along
and makes no secret of it. Reinhart's Coreference Rule is meant to throw light
on this well-known phenomenon.
Needless to say, variable binding is ruled out in the familiar way by
Condition B at SS. Coreference amounts to the following interpretation:
(12)
a.
she praises her to the sky
↓
↓
z
z
We must show that (12a) is not preempted under the Coreference Rule by any
other structure. A potential competitor with the right linguistic shape would be
(12b).
(12)
b.
she λ1[t1 praises her 1 to the sky]
↓
z
213
Irene Heim
If this doesn't qualify to preempt (12a), it can only be because its interpretation
is distinct. Is it?
Not if we just compare the propositions expressed. They are one and
the same, viz. that z praises z to the sky. So if we continue to interpret the
Coreference Rule in the way we did in the previous sections, it founders on this
example. Alternatively, we must look for a more suitable notion of
indistinguishable interpretation that will make it work right. The latter is, of
course, what Reinhart intends.
It is commonplace in the philosophical literature (especially on identity
statements) to distinguish the proposition expressed by an utterance from its
cognitive value . 9 When the utterance contains referring terms 10, the proposition
expressed depends only on their referents, but the cognitive value depends also
on the way these referents are presented. In the context of our example (12), for
instance, the person z presents herself to the interlocutors in two different
guises: First, they have a current visual impression of her, standing on the
platform over there and speaking. Second, they carry in their memory an entry
with various pieces of information about her, including that she is called
"Zelda".
Now when it comes to processing our sentence she praises her in (12),
what intuitively goes on seems to be this: Each of the two pronouns connects to
its referent z via one of these two guises. She, because of a perceived anaphoric
link to the subject this speaker of the preceding sentence, associates with the
visual impression; and her, through its link to the postcopular NP Zelda,
activates the memory entry. Therefore, the cognitive value of the sentence she
praises her to the sky for the hearer in this context is the proposition that
whoever causes the visual impression in question praises whomever the
pertinent memory entry represents. This is rather a different proposition from
the one the sentence expresses, viz. that z praises z.
More importantly for our present purposes, it is also different from the
cognitive value of the potential competitor in (12b). How so? Because (12b)
contains only one referring NP (she) and this will pick out its referent z via one
of the two guises salient in this context. Presumably this is the same one as for
the she in (12a), namely the visual impression. The cognitive value of the LF in
(12b) thereby comes to be the proposition that whoever causes this visual
impression praises herself—clearly a different proposition (with different
truthconditions) from the cognitive value of (12a) as described above.
The level of cognitive values thus seems to be more appropriate than
that of propositions expressed when it comes to distinguishing interpretations in
the sense of Reinhart's Coreference Rule. In this respect, our little diagrams
representing interpretations have been misleading or at least incomplete. An
utterance context for referring pronouns doesn't just supply these pronouns with
referents. Rather, it supplies them with guises, and these in turn happen to be
guises of something, namely the referents. The diagrams should thus contain an
9While some distinction along these lines is commonplace, many details and, of course,
the terminology vary from author to author. (The term "cognitive value", for instance,
comes form the introductory passage of Frege 1892, as translated by Max Black in Black
& Geach 1952.) For a thorough introduction and overview, see Haas-Spohn 1995.
10More accurately: directly referential terms in the sense of Kaplan 1989.
214
Anaphora and Semantic Interpretation
additional intermediate layer for the guises. (I use capital letters 'F', 'G', ... for
guises, with each letter standing for a unique guise and vice versa.) (12c)
replaces (12a), and (12d) replaces (12b).
(12)
c.
she praises her to the sky
↓
↓
F
G
↓
↓
z
z
d.
she λ1[t1 praises her 1 to the sky]
↓
F
↓
z
Semantically speaking, a guise is an individual concept, i.e. a function from
worlds to individuals. For instance, the visual impression alluded to above
(represented by 'F' in (12c,d)) can be viewed as assigning to each possible world
w the individual that it is an impression of in w. In the actual world of the
utterance, this is z, but in other worlds it may be someone else. The cognitive
value of (12c) is the proposition that is true in any w where F(w) praises G(w) to
the sky; that of (12d) is the proposition that is true in any w where F(w) praises
F(w) to the sky.
I will not formalize this any further, and in subsequent sections, I will
even fall back into the simplifying pretense that utterance contexts match
pronouns simply with referents. But before it's safe to do so, we should make
sure that the stories we told about (8) - (10) in the last section haven't collapsed
in the light of our refined notion of indistinguishable interpretation.
For instance, we took coreference between John and him to be ruled out
in (8) because (8d) preempted (8c). It did so, we said, because it expressed the
same proposition. But now we have seen that sameness of proposition
expressed is not a sufficient condition for indistinguishable interpretations. So
our argument re (8) is no longer conclusive. We should have established that
(8d) has the same cognitive value as (8c). Can this stronger argument be made?
Well, it can, if we bring out and exploit a tacit assumption about the example,
namely that it was meant to be judged either out of context, or in some sort of
run-of-the-mill context, say a conversation about John in his absence. In that
kind of ordinary setting, there wouldn't be multiple salient guises of j, but just
one (presumably the memory entry under his name), and the context would
assign that one to both the name and the pronoun in (8c). And in that case, the
cognitive values of (8c) and (8d) coincide, as desired. (Even if j happened to be
presented in two ways—say, he was visible in the distance during the
conversation about him—, this wouldn't suffice for the pronoun to automatically
link to him via a different guise than the name. For that to happen, the context
moreover has to contain appropriate clues that this is the intended
disambiguation.)
In short, we must qualify our earlier conclusion about (8): It doesn't
really follow from the grammar and the Coreference Rule alone that (8c) is
215
Irene Heim
unavailable; it only follows under additional assumptions about the context,
which imply that there is only one salient guise of j. But come to think of it, this
qualification is a good thing. It is a fact, after all, that even (8) could be used
with John and him coreferring if it were placed in an appropriately contrived
context.
There is much more to be said about the pragmatic conditions under
which contexts make two distinct guises of the same referent readily enough
available11, but let's move on.
2.4. Group (iv): when structured meaning matters
Evans (1980) emphasized a type of example that G&R likewise bring up in
illustration of the Coreference Rule. It will point us to yet another aspect of the
notion of indistinguishable interpretations. (The examples are again not exactly
Evans's or G&R's, but adaptations thereof to Condition B configurations.)
Consider the last clause of (13).
(13)
(You know what Mary, Sue and John have in common? Mary admires
John, Sue admires him, and) John admires him too.
Despite the Condition B environment, one sort of gets away with him referring
to John here. Apparently there is something about this particular preceding
discourse that makes it possible. What exactly is it and how does the
Coreference Rule predict it to matter?
The coreferential interpretation in question is (13a), and for some
reason it is not preempted by (13b).
(13)
a.
... and John admires him
↓
↓
j
j
b.
... and John λ1[t1 admires him1]
↓
j
11In Heim 1988:315 - 320, I proposed one concrete restriction: No context ever assigns
distinct but presupposedly coreferential guises to any pair of NP-occurrences.
(Definition: Guises F and G are presupposed to corefer in context c iff F(w) = G(w) for
every world w that conforms to the shared presuppositions of the discourse participants in
c.) This means that reference to the same object via two distinct guises is possible only
as long as the speaker still treats it as an open question whether indeed the same object is
behind these two guises. Once this is taken for granted (more accurately: presupposed in
the sense of Stalnaker 1979), only one guise is available. (This may be the result of
"collapsing" two previously available guises; in technical terms, the result of collapsing F
and G is F restricted to the set of worlds on which it coincides with G.) I still think this
proposal is defensible, though there are non-trivial issues to sort out (see, e.g. Landman
1986:104 - 105 for critical discussion). For the purposes of the present article, however, I
need not commit myself.
216
Anaphora and Semantic Interpretation
Why not? The propositions expressed are again identical, and there also don't
seem to be two different guises of j in this context that would lead to different
cognitive values. What then is the difference? It is, I think, the fact that two
different properties are predicated of j in (13a) and (13b): the property P of
admiring j in (13a), and the property Q of admiring oneself in (13b). In other
words, (13a) and (13b) express different structured propositions 12, <j,P> and
<j,Q>, even though they express the same unstructured proposition P(j) = Q(j) =
that j admires j.
Fair enough, but why does this suddenly matter? We could have
observed the same thing about all our earlier examples: there, too, the structured
propositions differed, even where we concluded that the interpretations were
indistinguishable. Evidently, differences in structured meaning do not always
matter. There has to be a special reason when they do. The special reason in
this case emerges when we analyze the preceding sentences: The speaker starts
out with a promise to tell what Mary, Sue, and John have in common. So we
expect him to mention a property that each of the three has. One particularly
direct and rhethorically effective way of doing this is to utter three predications
in which the same property is predicated of each of the three people in a row.
Now the first two predications, concerning Mary and Sue, were as follows:
(13)
c.
Mary admires John, Sue admires him, ...
↓
↓
↓
↓
m
j
s
j
Each of these predicates of its subject the property P of admiring j. Now if we
continue by (13a), we get a third predication of that same property. (13b), on
the other hand, would break the pattern and predicate a different property of j
than of the previous two people. It would still, of course, give us indirect
information about what the three have in common: we could determine the
shared property by a simple bit of deduction. But (13b) is more explicit, and
this, I submit, makes it beat out its competitor here and avoid being preempted.
A similar story applies to (14). This utterance, we are to imagine spoken by a
logic tutor.
(14)
Look, if everyone hates Oscar, then it surely follows that
Oscar (himself) hates him .
The intended interpretation is (14a).13
(14)
a.
if everyone hates Oscar, ... Oscar hates him
↓
↓
↓
o
o
o
Why isn't it preempted by (14b)?
12Structured propositions (and other types of structured meanings) have been put to
various uses in semantics; see Cresswell & Stechow 1982 for a recent example.
13Since subject quantifiers are interpretable in situ, I didn't bother to index and QR
everyone, but of course it wouldn't have hurt to do so.
217
Irene Heim
(14)
b.
if everyone hates Oscar, ... Oscar λ1[t1 hates him1]
↓
o
↓
o
Well, what the logic tutor is apparently trying to get across to the student is how
to apply the law of Universal Instantiation. (14a), as it happens, is a pure
illustration of that law: the predicates following everyone in the antecedent and
Oscar in the consequent denote the same property, that of hating Oscar. (14b),
by contrast, has the property of hating Oscar in the premise, but a different one,
that of hating oneself, in the conclusion. Of course, it is likewise a valid
inference. But its validity relies on more than just Universal Instantiation; it
collapses two inference steps (U.I. and λ-conversion). And that isn't optimal
didactic practice in this context.
So once again, a difference in structured meaning alone matters enough
to allow the interpretations (14a) and (14b) to count as distinct. But here as in
the previous example, this is due to very special circumstances: Logic teachers
have a professional duty to care not just about what proposition a sentence
conveys, but about how that proposition is built up from parts. For most
ordinary conversational purposes, however, the net message is all that counts,
and so we were right to disregard mere differences in structured meaning with
our earlier examples, and to disregard them again with most of the ones below.
2.5. Group (v): Lakoff's example14
Finally, why does one get away with utterances like (15), due to Lakoff (1972:
639)?
(15)
I dreamt that I was Brigitte Bardot and I kissed me.
The two underlined pronouns, being both first person, supposedly can't help but
corefer:
(15)
a.
... and I kissed me
↓
↓
g
g
('g' for George Lakoff, the speaker of (15).) But this interpretation ought to be
preempted by (15b).
(15)
b.
... and I λ1[t1 kissed me1]
↓
g
14This section is substantially changed from the previous version of this paper, partly in
response to questions raised by Higginbotham (p.c.) and Reinhart 1991b. Thanks to
Arnim von Stechow for reminding me of Stechow 1982. (McKay 1991 also looks
relevant, but I only received it when I was almost finished.)
218
Anaphora and Semantic Interpretation
Unless, again, we can argue that the latter has a distinct interpretation. That this
should be the case is made plausible by the observation that a minimally
different sentence, whose grammar forces the variable binding we see in (15b),
would actually describe a different dream:
(16)
I dreamt that I was Brigitte Bardot and I kissed myself.
(16) describes a dream about self-kissing, (15) doesn't. 15
Suggestive though this is, it is not so easy to show concretely how the
interpretations of (15a) and (15b) differ. The propositions expressed under the
reference assignments depicted in (15a,b) are certainly the same. The structured
meanings are different, but it is not evident why this should matter in this
context; at least there is no reason of the kind we found with (13) and (14)
above. Are the cognitive values distinct, then? For all we know so far, this
could only be if George Lakoff, as he utters (15), is somehow presented to his
audience in two separate guises, and that doesn't seem to be the case here either.
So where could the difference possibly lie?
To overcome this puzzle, we have to be a bit more precise on how such
examples are semantically interpreted. A suitable analysis of pronouns in
attitude reports, including an explicit application to (15), is found in Stechow
1982. Adapting his analysis (with some inessential alterations), let's begin by
positing a more articulate LF representation for attitude complements than we
took for granted in (15a,b). Following Quine 1956, believe, dream, and other
attitude verbs are logically 3-place predicates: the basic notion is for a subject to
believe (dream) something of something . The third argument (the "
res "argument) does not correspond to a surface constituent, but it is present at LF
and may be filled there by material moved out of the complement clause.16 For
instance, a simplified version of (15) (omitting the complement's first conjunct)
may have the following LFs, among others. (The corresponding SSs that these
should derive from in Reinhart's framework—all well-formed by the Binding
Conditions—appear in parentheses underneath.)
15Actually, the data are not quite so simple. The choice of the reflexive seems to be
compatible with both readings; at least this is my intuition about similar German
sentences:
(i)
Hans soll sich mal vorstellen, der Lehrer zu sein und sich/?ihn als Schüler zu
haben.
'Hans should imagine being the teacher and having himself/him as a student.'
The variant with the pronominal ihn is unambiguous as predicted by the analysis I will
sketch, though a little marginal (see below on what the marginality might be due to.) The
reflexive sich, however, also allows the pragmatically preferred reading, according to
which Hans imagines teaching Hans (rather than self-teaching). I have no account for
this reading. The analysis sketched in the text predicts only the self-teaching
interpretation here. Further research is required.
16Unfortunately, the syntactic distribution of
de re construals doesn't exhibit the
properties of movement. An "in situ" approach of the type that has proved successful for
association with focus would therefore be more appealing. The techniques for this are in
principle well worked out (see e.g. Rooth 1985) - except (to my knowledge) for the
interaction with variable-binding, which happens to be crucial in the present application.
219
Irene Heim
(15)
c.
I dreamt λ1[t1 kissed me] [ I ]
(SS: I dreamt I 1 kissed me )
d.
I dreamt λ1[I t1-ed me] [ kiss ]
(SS: I dreamt I kiss1ed me )
e.
I dreamt λ1,2[t1 kissed t2] [ I, me ]
(SS: I dreamt I 1 kissed me2 )
f.
I dreamt [ I kissed me ] [ ]
(SS: I dreamt I kissed me )
The "res-movement" that creates these LFs is like QR insofar as it involves λabstraction, but in some other respects operates quite differently: First, it isn't
Chomsky-adjunction, but substitution into a kind of argument position. Second,
it may affect phrases other than NPs (as in (d)). Third, it can apply multiply (as
in (e)). (If pure de dicto readings are to be covered as a special case, the res-slot
may also be left empty, as in (f).)
The semantic interpretation of the verb and its two internal arguments is
as follows: The left argument, denotes (as the λ-notation implies) a property of
n-tuples (n≥0, depending on the number of indices on the λ). The res-argument
denotes an n-tuple made up of the denotations of the phrases in it. The
interpretation of the verb is relative to a special contextual parameter, an n-tuple
of acquaintance relations in the sense of Lewis (1979). More precisely:
(17)
An LF of the form [α β γ [δ 1,...,δ n] ], where β is an attitude verb and n≥0,
requires an utterance context c which furnishes, for each i = 1, ..., n, an
acquaintance relation Dci.
To complete the analysis, take dream to mean 'believe in one's sleep'. In a given
context c, dream then denotes the following function f dream, c :
(18)
fdream, c (P)(<a 1,...,an>)(b) = 1 iff
(i)
b uniquely bears Dc1, ..., Dcn to <a1,...,an>, and
(ii)
b is asleep and self-ascribes the property of uniquely bearing
Dc1, ..., Dcn to some n-tuple of individuals satisfying P.
Now back to our example (15). Stechow proposes that (15) under the intended
reading has the LF-representation in (15c).17 The utterance context under
17It might actually be more accurate to assume (15e), with
both pronouns res-moved.
This would come to exactly the same reading if the second acquaintance relation supplied
by the context happened to be that relation which each individual bears to George Lakoff,
220
Anaphora and Semantic Interpretation
consideration (as before) assigns g as the referent to all three pronounoccurrences, and (by (17)) moreover supplies an acquaintance relation D. Here
is another picture to summarize these aspects of LF and context. (The pointer
under the verb indicates the contextually supplied acquaintance relation.)
(15)
g.
I dreamt λ1 [t1 kissed me] [ I ]
↓ ↓
↓
↓
g D
g
g
What is D? It is simply identity; this, at any rate, seems to be the choice that
yields the intuitively salient meaning. What the utterance asserts, then, is that g
self-ascribes in his sleep the property of kissing g. (The choice of identity for D
amounts to a so-called de se reading for the embedded subject; as Stechow
notes, this is usually the preferred reading for pronouns coreferential with the
higher subject.18)
Now, I think, we can finally see how G & R intend the Coreference
Rule to apply to this example. The me in (15g) is a referring pronoun that
corefers with the other two 1st-person pronouns in the sentence, in particular
with the embedded I that originates from a c-commanding position in its GC.
Why is this allowed?19 It would not be, (7) predicts, if (15g) meant exactly the
same thing as the following competitor with a bound variable:
(15)
h.
I dreamt λ1[t1 kissed me1] [ I ]
↓ ↓
↓
g D
g
In (15h), me has been replaced by a variable (me 1) A-bound by t1, the trace of I,
just as (7) demands. (The only difference compared to previous applications of
(7) is that the trace in question is a res-movement-trace, rather than a QR-trace).
(15h) has demonstrably different truthconditions from (15g): Given the same
value for D (identity), (15h) claims that g in his sleep self-ascribed the property
i.e. to that world-mate who is the actual Lakoff's counterpart by match-of-origins. It is
doubtful, however, that Lakoff knows enough about his origins to dream such a thing; see
Lewis 1984 for discussion. An advantage of (15e) over Stechow's choice (15c) would be
that the second pronoun could then be interpreted w.r.t. a similar, but probably more
realistic, acquaintance relation, say, the relation that x bears to y iff y is a world-mate of
x's who has the history and permanent characteristics that the actual George Lakoff
knows himself to have. But I will disregard this complication.
18For more discussion of de se readings and their status as special cases of
de re
readings, see Lewis 1979, Chierchia 1991, Higginbotham 1989, and Reinhart 1991b.
19Stechow also asks himself why (15) is not a Condition B violation, but his answer is
not quite sufficient for our present purposes. According to him, (15g) is okay because the
GC for me here is t 1 kissed me, which does not contain a coreferential c-commanding NP
(in fact, it contains no other referrring NP at all). But if this were good enough, why
couldn't we rescue every Condition B violation simply by QRing the offending
antecedent (e.g. as in John λ 1 [t1 saw him])? At any rate, it is not good enough for
Reinhart, whose theory we are assuming here. Under her assumptions, only looking at
(15g) itself is not enough to license it; we must also consider potential competing
structures.
221
Irene Heim
of self-kissing (the meaning earlier observed in sentence (16)). Hence (15h)
does not preempt (15g) and so (15g) is licensed.
3. Reference isn't special
Up to now I have merely tried to explain Reinhart's analysis. If I have gone
beyond plain repetition from G & R and her earlier publications, it was only to
flesh out details in the semantic analysis of certain examples, but not to add
anything that wasn't there yet, at least between the lines. In this section, I will
begin to disagree.
Let us look again at our fourth group of examples, the ones where
structured meaning mattered ((13) and (14)). Reinhart's account of them
captures an important intuition: These examples are licensed because of a
contextually important aspect of their meaning that would get lost if they were
replaced by their bound-variable counterparts. As such they call for a rule of the
kind of Reinhart's Coreference Rule, which essentially involves a comparison
between the meanings of two competing structures.
I still have a quibble, however: Is the phenomenon illustrated by (13)
and (14) really peculiar to referring NPs? Is it only referential pronouns that we
are sometimes allowed to use in unusual ways when a conversational purpose
justifies it? Come to think of it, such a limitation wouldn't be particularly
plausible to expect in an essentially pragmatic principle of this sort. And
indeed, once we start looking for the relevant examples, it isn't well supported
empirically either.
Recall, for instance, our logic tutor and his excuse for the coreferential
use of Oscar and him in (14). Once we let him get away with (14), are we really
going to put our foot down when he goes on as in (19)?
(19)
... And, of course, this doesn't just hold for Oscar, but for any arbitrary
man: If everyone hates a man, then that man himself hates him .
My point, of course, is that the last sentence in (19) is a donkey-sentence and the
two underlined NPs are donkey-anaphors, hence not referring terms. Under one
analysis of donkey-sentences, they are plainly bound variables, i.e., the LF
should be something like (19b), with a silent adverb of quantification equivalent
to a restricted universal quantifier (here abbreviated as "∀"). 20, 21
20Under an alternative (E-Type) approach to donkey-sentences, the donkey anaphors are
not plain bound variables, but descriptions of some kind, and the silent universal
quantification is not directly over men, but over cases. (See Neale 1990, Heim 1990, and
others for recent discussion.) Still, the donkey-anaphors are not referring terms under
that alternative either, because they would contain a bound case-variable, and so my main
point goes through all the same.
21It is not entirely obvious at this point how this LF is to be derived and licensed. One
revision to the present system that is surely required, once we bring in donkey-sentences,
is that indefinites on the one hand and demonstratives (and other complex definites) on
the other be allowed to count as variables, along with pronouns and anaphors. The
definition of variables also needs extending for donkey anaphors that are ordinary
pronouns, because these need not be A-bound. A suitable pair of added clauses might be:
222
Anaphora and Semantic Interpretation
(19)
a.
SS:
* if everyone hates [a man] 1,
[that man himself]1 hates him 1
b.
LF:
∀1 [ if [a man]1 [everyone hates t1] ]
[ [that man himself]1 hates him 1 ]
But this cannot be generated under Reinhart's assumptions: the derivation would
be blocked by Condition B at SS. There is no provision for licensing ill-formed
derivations like this, and the Coreference Rule is simply not pertinent here at all.
I take it that this is not satisfactory: (19) should be predicted to have pretty
much the same status as (14), with the choice of him over himself allowed for
the very same reason. Admittedly, (19) is (even) more contrived than (14) and
highly questionable from a pedagogical point of view. But it seems to me that
its added complexity is sufficient to account for what degradation we perceive
here. There isn't a qualitative contrast that would justify the special status of
referring pronouns that Reinhart's Coreference Rule in its present form grants
them.
And while you are in the appropriately contrived mind-set, consider a
bound variable variant of our other "structured meaning" example, (13):
(20)
Somebody said that what he had in common with his siblings was that his
sister admired him, his brother admired him, and he (himself) admired
him.
Again, it may not be the most natural English sentence, but the judgment isn't
such as to warrant a fundamental disparity between referential and bound
variable pronouns. My conclusion, therefore, is that Reinhart's Coreference
Rule captures the right intuition of what makes utterances like (13) and (14)
possible, but it should be made a little more general so that it covers (19) and
(20) as well.
Analogous points could be made about some of the other groups of
examples, but are more easily obscured there by murky technical details.
Recall, for instance, the only-example, whose intended interpretation is repeated
here:
(11)
a.
only he himself λ1[t1 pities him]
↓
↓
l
l
This, too, has bound-variable counterparts with much the same intuitive status.
For instance, (21) allows the reading in (21a).
(iv)
(v)
or on an indefinite in the restriction of a QAdv and A-bar-bound by it,
or on a definite in the nuclear scope of a QAdv and A-bar-bound by it.
223
Irene Heim
(21)
Every devil knows that only he himself pities him.
a.
∀x[devil(x) → know(x, that ∀y[y≠x → ~pity(y,x)])]
In Reinhart's system, this reading must have a derivation such as (21b,c).
(21)
b.
SS:
every devil1 knows that
[only [he himself]1]2 pities him1
c.
LF:
every devil λ1[t1 knows that
[ only [he himself]1 λ2[t2 pities him1] ] ]
Perhaps this is okay because no well-formedness constraint at either level rules
it out (in particular, the SS doesn't violate Condition B, because the adjoined
only blocks A-binding of him1 by [he himself]1). It is then not directly
problematic for Reinhart's theory. But isn't it a little strange if the explanation
for the acceptability of reading (21a) in (21) is so completely unlike the
explanation that was given for the acceptability of reading (11a) in (11)? The
latter involved comparison with potential preemptors under the Coreference
Rule, whereas the former relies entirely on considerations of syntactic wellformedness. We may suspect that a generalization is being missed here. It
would require some work to turn this suspicion into a real objection, but since I
won't let it carry the burden of my argument, I can afford to stop short of this
here.
The Lakoff-example likewise has bound-variable cousins, for instance
(22).
(22)
Not only I dreamt that I was Brigitte Bardot and I kissed me.
The relevant reading here is the one where (spoken by g) it says that there was
some x≠g such that x self-ascribed the property of kissing x (not to be confused
with the property of self-kissing). This seems no less acceptable than the
reading we have discussed for (15). But again, the explanation in Reinhart's
framework for why (22) allows such a reading cannot be anything like the
account given above for (15). Presumably, (22) has this reading because it
allows a derivation terminating in the following LF.
(22)
a.
not only I λ1[ t 1 dreamt that ... λ2[t2 kissed me1] [ I1 ] ]
It is not so clear at this point how such an LF is derived: what exactly is the SS
and why doesn't it violate Condition B? One way or another, these details must
be sorted out if the reading in question is to be generated. I will offer a concrete
suggestion below, but my present point doesn't depend on it. It is simply that,
whatever the details of an account of (22) in Reinhart's framework may turn out
to be, the Coreference Rule will not play any role in it. It couldn't, because there
is no coreference to be licensed (only a certain pattern of variable binding ). So
once again, what to the naive observer appears to be just a more complicated
224
Anaphora and Semantic Interpretation
instance of the same phenomenon receives a qualitatively different explanation - not, as it stands, a legitimate objection, but grounds for more suspicion.
The moral of this discussion seems to be that we should look for a more
general version of Reinhart's Coreference Rule, one that will cover bound
variable examples like (19) along with their simpler cousins involving
referential NPs, like (14). What we would like to replace (7) with is something
like the following:
(23)
Coreference-or-Cobinding Rule:
α cannot corefer or be cobound with β if an indistinguishable
interpretation can be generated by (indexing and moving β and) replacing
α with a variable A-bound by the trace of β.
This formulation presupposes a distinction between α and β being cobound on
the one hand and α being bound by (=cobound with the QR-trace of) β on the
other. (If "cobound" and "bound by" just meant the same thing, (23) would not
make sense, because then the cobound interpretation would be necessarily
indistinguishable from the one we are instructed to compare it to.) The
distinction is clear enough. For instance, [that man himself]1 and him1 are
cobound in our LF (19b) above, because each is bound by ∀1. If the latter were
to be bound by the former, the structure would have to look different, namely
like this:
(19)
c.
∀1 [ if [a man]1 [everyone hates t1] ]
[ [that man himself]1 λ2[t2 hates him 2] ]
The consequent clauses of (19c) and (19b) are logically equivalent, but they
differ in structured meaning, in just the way we have found to matter to the logic
tutor. So it is plausible that reading (19b) is available for (19) because of this
difference between it and (19c). (23), which predicts this, is on the right track.
The main point of this article is that (7) should be generalized to
something like (23), and having made that point, couldn't I stop here? I could if
it weren't for the following remaining loose ends and problems. First, I haven't
been able to be very precise yet about the treatment of the other types of
examples introduced in this section, i.e., (21) and (22). Second, (23) contains an
explicitly disjunctive formulation which it would be nicer to avoid. Third, even
with the generalized principle (23) in place of the former (7), the theory still
partitions the phenomena in a strange way: In coreference cases like (14), (23)
serves to license a certain interpretation for a grammatical derivation, but in
cobinding cases like (19) it acts to redeem an ungrammatical one. The actual
status of the examples does not warrant such a distinction; it would be better if
they were all grammatical, or all ungrammatical. Or, even better yet, if they
were all somewhere in between, which is what I will actually say below. I will
return to these points in section 5. The next section is devoted to an independent
criticism of Reinhart's system. The reason why I insert it at this juncture is that
it leads to some technical refinements that will be useful below.
225
Irene Heim
4. From coindexing to linking and codetermination
4.1. Bound variable pronouns that undergo QR
A technical difficulty arises when we ask what exactly happens when a bound
variable pronoun undergoes QR. Now I must first convince you that this
situation ever arises in the first place. To be sure, we have made QR completely
optional and unrestricted, so it would require a special stipulation to prevent it
from applying to bound variable pronouns. But do we really ever need to
exercise this option? The answer is 'yes': If we accept Reinhart's analysis of
ellipsis, there will be readings of English sentences that we can only generate by
QRing a bound variable pronoun.
This is not the place to launch into a detailed discussion of ellipsis. I
will just give a very brief exposition of Reinhart's approach and, for simplicity,
will concentrate entirely on Bare Argument ellipsis, setting aside any of the
additional complications that arise with the more common and colloquial VP
ellipsis. Consider a simple ellipsis structure:
(24)
I called John, and the teacher too.
(24) is ambiguous: the second conjunct can mean that the teacher called John,
or that I called the teacher. In the first reading, the "correspondent" of the
"remnant" the teacher is the subject I, in the second reading, the object John is
the correspondent.22 The basic idea, going back at least to Sag 1980, is that an
LF for the elliptical conjunct is derived by (a) QRing the correspondent in the
antecedent sentence, and (b) inserting a copy of the resulting λ−abstract next to
the remnant. Depending on the intended reading, this procedure yields (24a) or
(24b) for (24). (The copied material is in italics.)
(24)
a.
I λ1[t1 called John], and the teacher λ1[t 1 called John] too
b.
John λ1[I called t1], and the teacher λ1[I called t1] too
The semantic interpretation of these LFs is transparent.
In the two readings of example (24), the correspondents were a
referential pronoun and a proper name. But it is easy to construct similar
examples where the correspondent is a bound variable pronoun. A case in point
is the reading of (25) where it means that every boy said that I called both him
(the boy) and the teacher.
(25)
Every boy said that I called him, and the teacher too.
To derive the appropriate LF, we must QR the correspondent, which in this case
is the bound variable pronoun him. How exactly does this derivation proceed?
22The terminology of "correspondents" and "remnants" comes from the discussion of
Gapping in Pesetsky 1982.
226
Anaphora and Semantic Interpretation
First, what's the SS we start out from? Presumably, it has him coindexed with
every boy.
(25)
a.
SS:
every boy1 said that [I called him1, and the teacher too]
(The bracket just serves to indicate that the conjunction is in the lower clause.)
On the way to LF, both every boy 1 and him1 must QR, the former to bind the
pronoun, and the latter to create a constituent for ellipsis copying. So far, we
have assumed that QR applies in such a way as to shift the index on the moved
phrase over to the newly created λ next to it, so that the moved phrase itself
winds up unindexed. And this made sense in view of the fact that indices on
quantifying phrases like every boy (as well as indices on proper names and on
deictic pronouns) could make no semantic contribution anyway. However, if we
blindly applied QR to him 1 in exactly this same fashion, we get something
undesirable, namely (25b).
(25)
b.
every boy λ1[t1 said that
[him λ1[I called t1], and the teacher λ1[I called t1] too]]
The him has lost its index to the λ next to it and therefore is no longer bound by
(the λ next to) every boy. So (25b) cannot represent the intended reading.
It is not hard to think of ways to avoid this problem. The most obvious
and elegant option that comes to mind is to make the disappearance of the index
under QR simply optional: we are free to either retain the index of the QRed
phrase on both itself and the λ, or else—as before—to retain it only on the λ. In
most cases, we will effectively be forced to the second choice, because we
would otherwise end up with indexed NPs that don't qualify as variables under
(5). But nothing prevents us from retaining two copies of the index in the
special case where we are QRing a bound variable pronoun, and thus we can
derive an appropriate LF from the SS (25a), namely (25c).
(25)
c.
every boy λ1[t1 said that
[him1 λ1[I called t1], and the teacher λ1[I called t1] too]]
This expresses the intended meaning, and so we seem to have solved our
problem.
But wait, there is a complication: Consider the following slightly more
complex example:
(26)
Every boy said that [he called his mother and the teacher too].
This has many different readings, most of which are to be ignored here. We are
only interested in readings where (a) the correspondent of the the teacher is the
embedded subject he, and (b) both he and his are anaphorically related to every
boy. Still, there are two distinct readings that fit these specifications. The
difference comes out in the following two paraphrases for the elliptical conjunct:
227
Irene Heim
(26)
a.
... and the teacher called his own (the teacher's) mother.
b.
... and the teacher called his (the boy's) mother.
(26a) is a sloppy reading and (26b) a strict one. What SS and LF representations
are associated with each of these two readings?
Under our present assumptions, there is only one SS and LF available
to represent either reading: Because both he and his are to be bound variable
pronouns and anaphorically related to every boy, we can't but assign the
following indexing at SS:
(26)
c.
every boy1 said that he1 called his 1 mother and the teacher too
From there on, we have no real choices (not counting derivations that terminate
in uninterpretable LFs or leave the he unbound). Two applications of QR and
ellipsis copying yield (26d).
(26)
d.
every boy λ1[t1 said that
[he1 λ1[ t 1 called his 1 mother]
and the teacher λ1[ t1 called his1 mother] too]]
(26d) represents the sloppy meaning (26a). But for the other reading, the strict
one in (26b), we are left without any possible derivation.
4.2. Inner and outer indices
To remedy this limitation, I propose that we allow pronouns to be doubly
indexed at SS already. They can have an inner index that encodes what they are
bound by, and an additional outer index to encode what they in turn bind. (This
will be made more precise right below.) The inner and outer index need not be
the same.
Such dual indexing may look like a new-fangled notational contrivance,
but the distinctions it is meant to express are anything but new. Two
particularly important precedents are found in the PTQ fragment (Montague
1974) and in the Linking framework of Higginbotham 1983. Both of these
provide two different analyses for a sentence like (27) (a simplified version of
(26), without the elliptical conjunct).
(27)
Every boy said that he called his mother.
In PTQ, one option (call it (a)) is to build up a sentence with three occurrences
of the same free pronoun: he1 said that he1 called his1 mother, then use S14
with the operation F10,1 to quantify in every boy. In this derivation, both surface
pronouns are bound (=lose their subscripts) simultaneously in the last step.
Another option (b) would be to generate he2 called his 2 mother, then quantify
he1 into this to yield he1 called his mother, then build up further to he1 said that
he1 called his mother, and finally quantify in every boy with F10,1. This time,
228
Anaphora and Semantic Interpretation
the his was bound earlier in the derivation, whereas the he had a subscript until
the last step. 23 Of course, the meanings are equivalent, as one translation
reduces to the other by λ-conversion.
In Higginbotham's framework, (27) is generated with two different
linked structures:
(27)
a.
every boy said that he called his mother
↑↑__________|
|
|____________________|
b.
every boy said that he called his mother
↑_________|↑________|
These correspond respectively to the (a) and (b) derivations in PTQ. Given the
semantics for linked structures in Higginbotham 1987:125–127, 130–131, they
are again logically equivalent.
My double-indexing scheme mimics the two PTQ-derivations (a) and
(b) as follows:
(27)
c.
d.
SS:
[every boy]1 said that [he1] called [his1] mother
LF:
every boy λ1 [t 1 said that he 1 called his 1 mother]
SS:
[every boy]1 said that [he1]2 called [his2] mother
LF:
every boy λ1 [t 1 said that he 1 λ2 [t 2 called his2 mother]]
The LFs here and their intended semantics should be self-explanatory, but the
new SSs and SS-to-LF-changes call for some comments. Notice the distinction
between [αi] (inner index only), [α]i (outer index only), and [ αj]i (doubly
indexed). Regarding the operation of QR, I return to our original view that it
always works the same way. The trace retains the outer index of the moved
phrase, but the moved phrase itself transfers it to the λ and thereby loses it.
Bound variable pronouns are no exception. If an NP doesn't have an outer index
at SS, it just can't QR. On the other hand, if it does have one, it must QR, or else
23Of course, these are just two out of infinitely many derivations. For one thing, there
are infinitely many isomorphic derivations to each (a) and (b) where 1 and 2 are replaced
by different numbers. More interestingly, there are infinitely many additional derivations
not isomorphic to (a) or (b) which also yield the same meaning. Not all of these can also
be distinguished by means of my double indices. For instance, in PTQ we might build
he1 said that he1 called his 1 mother, then quantify in he2 with F 10,1, and then quantify in
every boy with F10,2. This derivation is not replicable in my system -- unless I were to
allow triple indices at SS, but I am not aware of any empirical motivation for this. There
are, however, some more types of PTQ-derivations besides (a) and (b) that do correspond
to distinct double-indices-representations; one of them will in fact become relevant right
below, see (26e) and (26g). (This footnote was prompted by questions raised by Barbara
Partee (p.c.). I am aware, of course, that I have not even begun to seriously explore the
expressive capacities of the proposed notation and its relation to alternatives in the
literature.)
229
Irene Heim
its outer index will fail to qualify on the following revised definition of
"variable" and the LF will be flitered out. (Note the change in clause (iii).)
(28)
Definition of "variable", revised:
An index is a variable only if it is
(i) on a λ, or
(ii) on a trace and bound by a λ, or
(iii) the inner index of a pronominal or anaphor and A-bound.
It follows that a bound variable pronoun that is to QR needs two indices: an
outer one in order to QR and an inner one in order to be bound.
Let's return to our problem with the strict-sloppy ambiguity in example
(26). Instead of the previous single option in (26c), we now have two choices
for representing (26) at SS, even when both pronouns are to be bound and the
subject is the correspondent:
(26)
e.
[every boy]1 said that [he1]2 called [his1] mother and the teacher
too
f.
[every boy]1 said that [he1]2 called [his2] mother and the teacher
too
(26f) is, of course, just like (27d) above; (26e) is like (27c) as regards the
indexing of his, but has an additional outer index on he (which we need to
enable it to QR, a prerequisite for ellipsis copying).24 Each of these derives a
unique LF:
(26)
g.
every boy λ1[t1 said that
[he1 λ2[ t 2 called his 1 mother]
and the teacher λ2[ t2 called his1 mother] too]]
h.
every boy λ1[t1 said that
[he1 λ2[ t 2 called his 2 mother]
and the teacher λ2[ t2 called his2 mother] too]]
And these correctly express the strict and sloppy readings respectively.
24Ignoring the elliptical conjunct, a PTQ derivation that corresponds to the indexing
choices in (26e) would proceed by building up he2 called his1 mother, quantifying in he1
by F10,2 to yield h e1 called his 1 mother, building up to he1 said that h e1 called his1
mother, and finally quantifying in every boy by F10,1. After the second step, this is like
derivation (a), and as in (a), both surface pronouns are bound simultaneously in the last
step.
230
Anaphora and Semantic Interpretation
4.3. The strong version of Condition B
We have now solved our initial technical problem about bound variable
pronouns undergoing QR, but we have yet to explore the repercussions of our
solution, especially for the operation of the Binding Conditions. With NPs
allowed to bear two distinct indices at once, there no longer is a unique obvious
notion of "coindexing," and thus of the Binding Conditions which rely on it.25
A variety of different coindexing-concepts are in principle definable, including
the following two:
(29)
Definition:
a.
b.
β is linked to α iff α's outer index = β's inner index.
α and β are colinked iff α's inner index = β's inner index.
The terminology is deliberately reminiscient of Higginbotham's linked
structures. It'll make sense to you if you glance at (27a,b) while applying the
definitions to (27c,d). In fact, linking as defined in (29a) has exactly the
properties of Higginbotham's linking relation; for instance, unlike coindexing it
is neither symmetric nor transitive.
Consider now Condition A. In the version used so far, it requires that
an anaphor be "coindexed with" a c-commanding A-position in its GC. How
should we reinterpret it in the present setting? Higginbotham proposes to
replace "coindexed with" by "linked to":
(30)
Condition A, new version:
An anaphor is linked to a c-commanding A-position in its GC.
(30) predicts that an SS such as (31a) is ill-formed.
(31)
a.
* [he1]2 cut [himself 1]
The reflexive, though in some sense "coindexed" with the subject, is not linked
to it in the sense of definition (29a). What (30) requires instead is the kind of
indexing shown in (31b).
(31b) [he1]2 cut [himself 2]
The semantic import of this prediction is that a reflexive must really be bound
within its GC, not just have a cobound antecedent there. To see this, look at the
LFs that (31a,b) give rise to:
25The reference to "A-binding", and hence to "coindexing", in LF-conditions like (28)
remains unproblematic, since all indices are single by the time we reach a well-formed
and interpretable LF.
231
Irene Heim
(31)
c.
from *(31a):
he1 λ2[t2 cut himself1]
d.
from (31b):
he1 λ2[t2 cut himself2]
The reflexive is free in (31c) and bound in (31d). Neither of these, of course, is
a well-formed LF on its own on our current assumptions, because of the free
index on he1 . But they might be part of larger well-formed LFs such as (32a,b).
(32)
a.
every boy λ1[t1 said that he1 λ2[t2 cut himself1]]
b.
every boy λ1[t1 said that he1 λ2[t2 cut himself2]]
These two happen to be logically equivalent, so we cannot observe right here
whether the prediction that (32a) comes from an ungrammatical SS is borne out.
But if we add an elliptical conjunct, the difference becomes manifest: (30)
predicts, correctly, I assume26, that (33)—unlike (26)—allows only a sloppy
reading.
(33)
[every boy]1 said that [he1]2 cut [himself *1/2] and the teacher too.
Now let's turn to Condition B. This used to require that a pronominal not be
"coindexed with" any c-commanding A-position in its GC. Should we again
replace "coindexed with" by "linked to"? If we have learned the lessons of
Partee & Bach (1984)27 or Higginbotham (1983; 1985), we know better than
that. Consider the following potential derivation for sentence (34):
(34)
every boy said that he called him
a.
SS:
[every boy]1 said that [he1]2 called [him1]
26More precisely: if (33) allows a strict reading as well, it does so no more easily than
(i).
(i)
John cut himself, and the teacher too.
Whatever relaxation of Reinhart's assumptions (here left untouched) about reflexives and
ellipsis will accommodate a strict reading in (i) should do so for (33). I don't mean to
dismiss the issue of strict readings with reflexives or diminish its potential significance
for the goals of this paper, but one thing that seems fairly clear is that we don't want a
theory that draws the dividing line between (33) and (i).
27I am indebted to Barbara Partee for reminding me of this important reference. It may
be the earliest attempt to formulate Binding Conditions A and B within a framework
which (due to its kinship with PTQ, in this case) captures semantic distinctions that turn
on the difference between "cobound" and "bound by", such as the strict/sloppy ambiguity
in (26). Just as in Higginbotham's framework, the need to strengthen Condition B arises
as an immediate consequence of the system's ability to represent this semantic distinction.
Partee & Bach present the issue somewhat differently than I have done here: Their
treatment of ellipsis (which focusses on VP Ellipsis rather than Bare Argument Ellipsis)
does not rely on quantifying-in of pronouns. But the Derived VP Rule that they employ
instead has a similar effect of multiplying the possible derivations for sentences with two
pronouns (like (34)). A detailed comparison between their proposals and those of
Higginbotham or the present article must await another occasion.
232
Anaphora and Semantic Interpretation
b,
LF:
every boy λ1[t1 said that he1 λ2[t2 called him1]]
(34b) is logically equivalent to every boy λ1[t 1 said that he1 called him1], a
reading that sentence (34) clearly lacks and that English can only express with
the reflexive: 'every boy said that he called himself'. So this derivation should
be ruled out. If Condition B merely prohibits linking, it will not block (34a):
[him1] is not, after all, linked to [he1]2 (nor is there anything else we know of
wrong with the derivation). [he1]2 and [him1] are, however, colinked (as
defined in (29b)). So if we make Condition B sensitive to both linking and
colinking, we would seem to solve our problem. This suggests the following
formulation:
(35)
Condition B, preliminary new version:
A pronominal is neither linked to nor colinked with any c-commanding
A-position in its GC.
A noteworthy implication of such a conception of Condition B is that many
indexing patterns cannot be realized with any lexical item at all. For instance,
[he1] called [α1] is ill-formed however we spell out α, in particular for α = him
as well as α = himself.
(35) is adequate to deal with (34), but it is known to be still too weak
for similar examples with somewhat more complex indexing patterns, for
instance (36).28
(36)
* [every boy]1 said [he1]2 knew [he1] called [him2]
The LF that (36) gives rise to has the meaning of 'every boy said he knew he
called himself', so this better not be a well-formed SS. Treading again in
Higginbotham's footsteps 29, I therefore define an even more inclusive relation
of which linking and colinking are two, but not the only, special cases.
(37)
Definition30:
28This too was already seen clearly by Partee and Bach (1984). See especially their
section 6.3, pp. 124 - 25, where they comment on the sentence Every man talked to
himself by himself about a book that he gave to him and the need for a condition that is
sensitive to global properties of the derivation.
29See especially the motivation for the "transitivity condition" in Higginbotham 1983:
406 and the refined conditions that replace it in 1985: 570 - 575 and 1991: 8 - 11. The
following is an attempt to replicate his proposal that two elements in a Condition B
configuration "cannot be determined by the structure in which they occur to share a
value" (1985: 575).
30This is an equivalence relation. In the present system, it happens to be coextensive
with the reflexive, symmetrical, and transitive closure of the linked-to-relation. It will
differ, however, after we reintroduce free variables in section 5.2 below. Two NPs with
233
Irene Heim
α and β are codetermined iff
(i) α = β, or
(ii) either one of α or β is linked to the other, or
(iii) α and β are colinked, or
(iv) for some γ, α and γ are codetermined and so are γ and β.
(38)
Condition B, final new version:
A pronominal is not codetermined with any c-commanding A-position in
its GC.
I leave it to the reader to check this against (36) and other even fancier cases
suggested by the literature. For the remainder of this article, it is sufficient to
remember that codetermination includes linking and colinking.
To sum up section 4, I started out with the question of how QR applies
to bound variable pronouns and ended up with an endorsement of
Higginbotham's Linking Theory, in particular of its non-parallel conceptions of
Binding Conditions A and B. The overall indexing system appears more
complex now, but it may be seen as a conservative extension of Reinhart's ideas.
Notice in particular that, throughout my revisions, I have adhered to her
assumption that there is a perfect correspondence between syntactic binding in
the sense of Condition A and semantic binding (variable binding at LF). The
newly introduced double indices serve, in effect, to express distinctions at SS
that otherwise could only be brought out at LF. And the motivation for
expressing them there is (as throughout Reinhart's work) the fact that syntactic
Binding Conditions, which apply at SS, are sensitive to those distinctions.
5. Revised proposal
5.1. The Exceptional Coindexing Rule
We are now ready to return to the central issue of this study, namely the status
of the Coreference Rule. In the section 3 titled "reference isn't special", I
suggested that whatever will survive of this rule should pertain not just to
coreference possibilities but also to the possibility of cobinding two bound
variable NPs. Consider again the derivation I gave above for the donkey
sentence version of the logic tutor's utterance and which violated (the original
version of) Condition B:
(19)
a.
SS:
* if everyone hates [a man] 1,
[that man himself]1 hates him 1
b.
LF:
∀1 [ if [a man]1 [everyone hates t1] ]
matching free inner indices will then count as colinked by (29b), and hence as
codetermined by (37iii), and this is intended.
234
Anaphora and Semantic Interpretation
[ [that man himself]1 hates him 1 ]
From our current perspective, the question arises whether the index on that man
himself is supposed to be an inner or an outer index. Given that it encodes what
this bound variable demonstrative is bound by, it makes sense to consider it an
inner index, as in (19d,e).31
(19)
d.
SS:
* if everyone hates [a man1],
[ that man himself1] hates [him1]
e.
LF:
∀1 [ if a man1 [everyone hates t1] ]
[ that man himself1 hates him 1 ]
This derivation, at any rate, seems appropriate to the intended reading. Notice
how (19e) illustrates Universal Instantiation: under any given assignment, the
predicates following everyone in the antecedent and that man himself in the
consequent denote the same property.
Our worry above was that (19a) was excluded by Condition B. (19d) is
still excluded by the new version of Condition B: [that man himself 1] and
[him1] are colinked. If we say nothing further, we are no better off than before.
Suppose, however, we exploit the refined indexing system to draw a distinction
between different types of Condition B violations: linking violations and mere
codetermination violations. In the former, a pronominal is linked to a ccommanding A-position in its GC; in the latter, it is codetermined with one, but
not linked to any. Notice that [him1] in (19d) is codetermined with [that man
himself1], but not linked to it, so this is a mere codetermination violation. I
speculate that this is significant. The general hypothesis is as follows: Linking
violations of Condition B are disallowed under any circumstances. Mere
codetermination violations are normally disallowed as well, and perhaps are
never 100% perfect, but they can be much improved under suitable discourse
conditions. Specifically, what makes (19) viable could be a principle like the
following.
(39)
Exceptional Coindexing Rule:
A pronominal α is (marginally) allowed (at SS) to be codetermined with a
c-commanding A-position β in its GC when the interpretation thus
obtained needs to be distinguished from the one that would result if (β
31The pertinent clauses in the definition of variables (cf. footnote 21) are accordingly
reinterpreted as follows:
(iv)
or the inner index of an indefinite in the restriction of a QAdv and A-bar-bound
by it,
(v)
or the inner index of a definite in the nuclear scope of a QAdv and A-bar-bound
by it.
235
Irene Heim
were given an outer index and moved and) α were replaced by a variable
A-bound by the trace of β.
This is obviously a very close relative of Reinhart's Coreference Rule. Not only
does it express her basic insight that we may sometimes say things we otherwise
couldn't because that's the best way to get our communicative point across; it
also retains almost all the concrete wording of (7). But one thing is different:
(39) is about indices, not about referents.
(39) applies to sanction (19d,e), on the grounds that the meaning of the
competing structure in (19f) would not have served the logic tutor's special
communicative purpose:
(19)
f.
∀1 [ if a man1 [everyone hates t1] ]
[ that man himself1 λ2[t2 hates him 2] ]
Though logically equivalent, the structured meaning of (19f) fails to illustrate
pure Universal Instantiation.
Now it would be rather dissatisfying to have both the Coreference Rule
(7) and the Exceptional Coindexing Rule (39) side by side in our system. They
clearly concern different instances of the same phenomenon and we would like
to collapse them somehow. But how is this possible? Under the present
assumptions about syntactic representation which we took from Reinhart, there
is a fundamental difference between (7) and (39): (39) deals with licensing
certain indexing patterns, but (7) doesn't deal with any aspect of syntactic
representation at all, but directly with interpretive choices. The reason for this
is, of course, Reinhart's view that coreference relations, as opposed to variable
binding relations, have no syntactic representation. If we want to unify (7) and
(39), it seems advisable to give up this radical bifurcation. In particular, if we
want to subsume the cases now dealt with by (7) under the scope of (39), we are
driven back to the more conventional view of coindexing as a representational
device that does double duty for coreference and variable binding. This is the
line that I will pursue here, and it requires a couple of revisions in the basic
syntactic and semantic system.
5.2. Free variables readmitted
If referring NPs are to bear indices and indices are semantically variables, then
referring NPs correspond to free variables. Free variables were systematically
excluded in Reinhart's system, so our first task is to permit free variables under
the well-formedness conditions for LFs. The current definition of "variable"
leaves no room for them, so we have to extend it.
(40)
Definition of "variable", revised again:
An index is a variable only if it is
(i) on a λ, or
(ii) on a trace and bound by a λ, or
236
Anaphora and Semantic Interpretation
(iii) the inner index of a pronominal or anaphor and A-bound, or
(iv) the inner index of a definite NP and free.
Second, we must provide these free variables with some kind of semantic
interpretation. Let's at first restrict our attention to the special case where we are
dealing with free pronouns. These are, of course, the deictic pronouns (formerly
unindexed) and they get their reference assigned by the context. Since we want
to ensure that coindexed free pronouns are ipso facto coreferential, we should
now think of the context as supplying a referent per index (rather than, as before,
per pronoun occurrence). And mindful of our discussion of cognitive value in
section 2.3, we assume that it is not really a referent that the context supplies,
but a guise. So we add the following principle:
(41)
Reference Assignment (1st version):
An LF φ is interpretable in an utterance context c only if
for each index i free in φ, c furnishes a guise Fci.
In the compositional semantic interpretation of the utterance, these contextually
supplied guises provide values for the free variables. When we are calculating
the proposition expressed by an LF φ in a given context c, we assign to each free
αi the individual F ci(w c ) that Fci picks out in the utterance world wc . (To obtain
the cognitive value, we first have to calculate the propositional concept that
maps each world w in the context set of c to the proposition φ would express if
uttered in w, then diagonalize this; see Stalnaker 1979.)
As it stands, (41) ensures that coindexed free pronouns corefer. In fact,
it ensures something even stronger, namely that they pick out the same referent
under the same guise. How about two free pronouns with
different indices?
Could they still receive the same referent? Could they be assigned even the
same guise? Nothing so far prevents it, and as far as mere coreference goes,
that's just fine. But sameness of guises is something we have to disallow, for
reasons that will quickly become clear. So we need an amendment:
(42)
Reference Assignment (2nd version):
An LF φ is interpretable in an utterance context c only if c furnishes a
distinct guise F ci for each index i free in φ.
With this much of the new system in place, the coverage of our Exceptional
Coindexing Rule is no longer limited to bound variable examples like (19), but
includes analogous cases with two coreferential pronouns, such as the following
variant of (14).
(43)
Look, if everyone hates Oscar, then it surely follows that he himself hates
him.
Suppose the intended interpretation of this has the following derivation:
237
Irene Heim
(43)
a.
SS:
if everyone hates Oscar, ... [he himself2] hates [him2]
b.
LF:
if everyone hates Oscar, ... he himself2 hates him2
The utterance context c for (43b) is of course meant to be such that Fc2(w c ) =
Oscar. (Never mind what the complete appropriate guise of Oscar is.) So the
consequent clause of (43b) in context c expresses the structured proposition that
consists of Oscar and the property of hating Oscar. Now (43a) is a Condition B
violation because [he himself 2] is in the GC of and codeterrmined with [him2].
But (39) allows it anyway because this is a context where structured meaning
matters, and the alternative (43c) would express a different structured
proposition in c, namely the one consisting of Oscar and the property of selfhating.
(43)
c.
... he himself2 λ3[t3 hates him3]
So far, so good. Our account of this example is not complete, however, without
a demonstration that (43a,b) is essentially the only derivation under which it gets
the intended meaning (apart from trivial alphabetic variants, of course). In
particular, we have to watch out for the following two potential rivals, in which
he himself has either received no index at all, or a different one from him.
(43)
(43)
d.
SS:
if everyone hates Oscar, ... [he himself] hates [him 2]
e.
LF:
if everyone hates Oscar, ... he himself hates him 2
f.
SS:
if everyone hates Oscar, ... [he himself4] hates [him2]
g.
LF:
if everyone hates Oscar, ... he himself4 hates him 2
Neither of these are even prima face Condition B violations, so any
interpretations they represent should be fine quite regardless of the special
discourse conditions that are required for clearance by the Exceptional
Coindexing Rule. They better not be alternate representations of our intended
interpretation, then, because that way we would lose our empirical prediction
about (43). In other words, we would lose Reinhart's insight that this particular
interpretation of this English sentence stands and falls with the presence of a
context in which there is some special reason for structured meaning to matter.
As regards (43d,e), I will simply stipulate it away by a provision that
pronouns bear an (inner) index as a lexical property (see right below). (43f,g),
on the other hand, must certainly be allowed as a well-formed derivation for
some reading of (43), for example one where he himself refers to Fred. But can
it be set in a context c where it so happens that Fc4(w c ) = F c2(w c ) = Oscar?
Well yes, it could so happen —for instance if we have been talking about a
certain "Fred", he himself4 connects anaphorically to an occurrence of that name
Fred, and then Fred and Oscar turn out to be the same person. But in this case,
F c4 ≠ Fc2 (even though they concide on wc ), and that's a different interpretation
from the one we have been contemplating all along (and one that should be
predicted okay without special excuses about the importance of structured
238
Anaphora and Semantic Interpretation
meaning!). What our Reference Assignment principle doesn't allow is for he
himself4 and him2 in (43g) to pick out Oscar through the same guise. (This was
the purpose of the amendment in the 2nd version.) And this, I take it, suffices to
disqualify (43f,g) as an alternative representation of the interpretation we
associated with (43a,b). The logic tutorial setting that we have been imagining
just doesn't provide two separate guises of Oscar (much less appropriate clues to
determine which pronoun goes with which guise).
So much for (43). But of course we also need to cover the original
(14), with the proper name Oscar (himself) rather than a pronoun in the subject
position. In order to do this, we must allow the name to bear an (inner) index
and that index must represent the name's reference. What's more, we must not
just allow, but force , names to bear indices. This is for the same reason that we
wanted to exclude (43d,e) above: If we let a name be interpretable without the
benefit of an index, we couldn't read off coreference relations between that name
and some other NP from the syntactic representation. That way, we could
always "sneak by" Condition B and the Exceptional Coindexing Rule by
suppressing potentially offending indices and still get equivalent readings. And
then we would still need a separate Coreference Rule of Reinhart's kind to block
those.
So we must go all the way and make sure that there is no reference
without an index. Specifically, I stipulate that all pronouns, proper names, and
other definites bear an inner index as part of their lexical entry and thus come
automatically indexed when they are inserted in any syntactic structure. (The
assignment of outer indices may be left free and optional as before.)
Free indices still are assigned referents by the utterance context, but the
NP's lexical content must now be made to constrain this assignment. Actually,
to some extent this happens even with pronouns, which normally have gender,
number, and person features that limit their choice of referent. Proper names are
just a more extreme case of this phenomenon: he fits any male person or animal,
John is specialized to those named "John". Since that's a common name, there is
still considerable disambiguating work left to be done by the usual pragmatic
factors; but the formulation below is also meant to cover the limiting case of a
name unique enough to render the context's role trivial.
(44)
Reference Assignment (3rd version):
An LF φ is interpretable in an utterance context c only if c furnishes a
distinct guise Fci for each index i free in φ
such that it is presupposed in c that Fci picks out an individual
which fits the features and lexical content of any NP indexed i occurring
in φ.
(44) is a bit vague, but that could be worked out.32
There is now no significant difference between (43) and the original
(14). On our current assumptions, the latter's SS and LF are as follows:
32For some pertinent discussion, especially regarding definite descriptions, see Heim
1988.
239
Irene Heim
(14)
c.
SS: if everyone hates [Oscar2], ... [Oscar2] hates [him2]
d.
LF: if everyone hates Oscar2, ... Oscar2 hates him2
The context must match the free index 2 with some guise that is known to be of
something named "Oscar" and a male human or animal. (If the second
requirement, stemming from the features of the pronoun, is redundant, given that
"Oscar" is a familiar name for males, so much the better.) Presumably, this
guise is the memory entry under the name "Oscar".33 The Exceptional
Coindexing Rule applies to license this interpretation just as we showed above
for (43a,b), and (for the reasons we went over there) we needn't worry about
alternate derivations in which Oscar and him weren't coindexed.
5.3. Other examples in the revised analysis
In arguing that the Exceptional Coindexing Rule takes over the work of
Reinhart's Coreference Rule, I have looked at only one type of example so far,
the kind where structured meaning matters. What about the other types initially
adduced in illustration of the Coreference Rule? Let's run through the list one
more time. We will see that the revised account parallels G & R's original one
for the only and Lakoff examples. On the other hand, due to the fact that (39)
doesn't apply unless there is a Condition B violation in the first place, the basic
cases and debates about identity are now licensed independently of (39), simply
because they have fully grammatical representations.
5.3.1. Basic cases
Here is our initial trio again:
(8)
John saw him .
(9)
John saw his mother.
(10)
His mother saw John.
We can be brief on (8): A bound variable interpretation is out as before, as it
would be a linking violation of Condition B. A corefential interpretation, with
the common referent picked out under the same guise, would have to be
represented by matching inner indices on John and him and thus would be a
mere codetermination violation of B. It is therefore also out, except in special
contexts where structured meaning matters. Finally, coreference via distinct
guises is okay, but requires the sort of context where distinct guises are readily
available.
With (9), Condition B does not apply since John is beyond the GC of
his. Both linking and colinking are therefore permitted. The former represents
the bound variable reading, the latter the coreferential one. Here we see the
benefit of restricting (39) to prima facie ill-formed structures (as opposed to
33In the logic tutorial setting, there may be no reference to any particular real-life Oscar.
But let's abstract away from this bit of make-believe.
240
Anaphora and Semantic Interpretation
haveing it apply everywhere, as did (7) and (23)). As G & R point out in their
footnote 13, there is reason to believe that (10) is inherently ambiguous.
Otherwise one would have to say that it allows a coreferential interpretation in
continuations like John saw his mother, and so did Bill only because this yields
a manifestly distinct interpretation for the continuation. But this is problematic
for two reasons. First, as noted in Lasnik 1989:166, it doesn't straightforwardly
cover the case where the elliptical continuation is uttered by a different speaker.
Second, there is independent reason to doubt that the Coreference Rule "looks
ahead" that far in the discourse: if it did, John saw him, and Bill did too, should
likewise permit coreference between John and him. G & R show that these
problems disappear if the Coreference Rule is restricted to prima facie Condition
B violations, and I have taken up a version of that idea in my (39).34
In (10), variable binding continues to be out for the same reason as
before: Locally A-bar-bound pronouns still fall through the cracks of the
definition of "variable". But nothing whatsoever interferes with coreference,
which would be represented by matching inner indices on his and John. (Again,
I mean coreference via a common guise.)
5.3.2. Identity debates
Let's turn to example (12).
(12)
A:
Is this speaker Zelda?
B:
How can you doubt it? She praises her to the sky. No competing
candidate would do that.
In a way, I have already said everything I have to say about this one. Without
being mentioned explicitly at that point, this type of example was my guide in
setting up the Reference Assignment principle in such a way that it imposes a
one-to-one match of indices and guises (and not of indices and actual referents).
So whether or not party B turns out to be right and the speaker is Zelda, the
syntactic representation of the pertinent reading of this example will have
distinct inner indices on she and her, and so, once again, Condition B doesn't
even get a shot at it.
Here the present analysis actually takes a different view than G & R,
even in the version of their footnote 13, which restricts (7) to prima facie
Condition B violations. (12) does qualify as such a prima facie violation in their
sense, but it isn't a violation of any kind in the present system. This is welcome
insofar as this type of case has none of the marginal flavor of the other types (as
G & R also concede in their footnote 10).
5.3.3. only examples
Here is the example about Lucifer once more:
(11)
(Everybody hates Lucifer.) Only he himself pities him .
34G & R actually don't commit themselves to this particular solution, they just spell it out
as one possibility among others.
241
Irene Heim
Recall that there was a murky point in G & R's original account of this example,
a kind of equivocation between "the trace of only he himself " and "the trace of
he himself". Let us see what shape this takes in the present system. The
derivation of (11) on the intended reading ought to be (11c,d).
(11)
c.
SS:
[only [he himself1]] 2 pities [him 1]
d.
LF:
only [he himself1] λ2 [t 2 pities him1]
As indicated, 1 in (11c) is an inner index on the smaller NP, and 2 an outer
index on the larger NP. If this is indeed the situation, then (11c) is not in
violation of Condition B, not even a codetermination violation: the larger NP is
in a c-commanding A-position w.r.t. [him1] but doesn't share any index with it,
and the smaller one is colinked with [him1] but not in a c-commanding Aposition.
If we are only concerned with allowing this interpretation of (11), we
can live with this result. But the intuitive status of this type of example is like
that of others we have classified as weak Condition B violations in need of
sanctioning by the Exceptional Coindexing Rule. So a better prediction would
be that the subject and object in the SS of (11) do count as codetermined, and
that therefore (39) has to step in to license (11d) on the grounds that its truthconditions differ from those of (11e):
(11)
e.
only [he himself 1] λ2 [t 2 pities him2]
At this point, our options are not unlike G & R's: For instance, we can just
stipulate that adjunction structures of this sort count as one NP rather than two.
(This implies, among other things, that such structures never have room for
more than a total of two indices; they couldn't have up to four, as we'd expect in
a genuine instance of two nested NPs.) The indices 1 and 2 on he himself and
only he himself then just are the inner and outer index of this one NP, which
thereby qualifies as colinked with the object. As it stands, this is just as hokey
as the equivocation implicit in G & R's original story. But perhaps it can be
improved upon when future research gives us a deeper understanding of both the
syntax and the semantics of double indexing.35
35In a related type of example that is likely to be relevant in this connection,
only does
not form a constituent with the focus it associates with:
(i)
I could only see that [he himself]F pities him.
As shown by Rooth (1985: 72 - 72, 76 - 79), an index on a focussed phrase may be inside
or outside the focus-marker, with non-equivalent semantic results. In the present setting,
we face the question whether the phrase including the F and the one that excludes it
should be treated as two separate, nested NPs (each with potentially an inner and an outer
index), or whether there is just one NP, whose inner index is interpreted below the F and
whose outer index above it. The latter choice is preferable in light of the fact that the
relevant coreferential readings for such sentences have the same intuitive status as
analogous cases in which only and its focus form a constituent. -- The discussion of
anaphora with focussed antecedents in Higginbotham 1989 also points in this direction.
242
Anaphora and Semantic Interpretation
In other words, our analysis of (11) remains essentially
Reinhart's—except that it now extends automatically to (21), which receives the
derivation below:
(21)
d.
SS:
[every devil]1 knows that
[only [he himself1]] 2 pities [him 1]
e.
LF:
every devil λ1[t1 knows that
[ only [he himself1] λ2[t2 pities him1] ] ]
The fine-structure of the only-NP in (21d) is whatever it is in (11c), and thus
(11) and (21) are alike from the standpoint of both Condition B and the
Exceptional Coindexing Rule. The potential preempting alternative to (21e),
which however has different truth-conditions, is (21f).
(21)
f.
every devil λ1[t1 knows that
[ only [he himself1] λ2[t2 pities him2] ] ]
5.3.4. Lakoff's example
G & R's account of (15) likewise carries over in full and now generalizes to (22).
The derivations speak for themselves.
(15)
(22)
i.
SS:
[I2] dreamt [I2]1 kissed [me 2]
j.
LF:
I2 dreamt λ1 [ t 1 kissed me2 ] [ I2 ]
k.
potential preemptor: I2 dreamt λ1 [ t 1 kissed me1 ] [ I2 ]
b.
SS:
[not only [I3]] 2 dreamt [I2]1 kissed [me 2]
c.
LF:
not only [I3] λ2 [t 2 dreamt λ1 [ t 1 kissed me2 ] [ I2 ] ]
d.
potential preemptor:
not only [I3] λ2 [t 2 dreamt λ1 [ t 1 kissed me1 ] [ I2 ] ]
6. Concluding remark
I have pointed out plenty of loose ends along the way. Let me just note one
more respect in which the present account might need revision. My proposal, as
it stands, actually has two separate parts. One claim is that there are two
different types of Condition B violation, and that one of these types, the mere
243
Irene Heim
codetermination violations, can be rescued under suitable discourse conditions.
The second claim specifies more concretely what these discourse conditions are;
in close agreement with Reinhart, it implies that an interpretive difference
between the LF to be licensed and its binding-alternative is a sufficient licensing
factor.
Now it could conceivably turn out that the first claim is right, but the
second is wrong. Specifically, there may be additional discourse factors, besides
or even instead of the one that Reinhart identified. For instance, Evans (1980)
and others have conjectured that prima facie Condition B violations improve
when the previous discourse provides an antecedent that both potentially
offending NPs can be anaphorically linked to. G & R correctly observe that this
is not a sufficient condition for such violations to become acceptable. But a
possibility that has not been explicitly considered so far is that it might be a
necessary factor in addition to the requirement stated in (7) or (39). There is
something to be said for this possibility. All the relevant examples I have used
in this article actually happen to have this property; see, in particular, (11), (14),
(15), and their variants, all of which contained a suitable common antecedent for
the two NPs in the Condition B configuration. As Hans Kamp (p.c.) pointed out
to me, this appears to be crucial. For instance, if we have not been talking about
Lucifer before, an utterance of Only Lucifer pities him (or of Only he pities him,
using the first pronoun with a demonstrative reference to Lucifer) is quite bad
under a coreferential reading. So the condition identified by Reinhart and
carried over into my Exceptional Coindexing Rule is not by itself sufficient, and
a common antecedent seems to be required on top of it. This needs further
exploration.
References
Black, Max and Peter Geach eds. (1952) Translations from the Philosophical Writings of
Gottlob Frege, Oxford: Blackwell.
Chierchia, Gennaro (1991) "Anaphors and Attitudes De Se," in R. Bartsch, J. van
Benthem, P. van Emde Boas eds. Language in Context, Dordrecht: Foris.
Cresswell, Maxwell and Arnim Stechow (1982) "De Re Belief Generalized," Linguistics
and Philosophy 5.4, 503 -535.
Dalrymple, Mary, Stuart Schieber, Fernando Pereira (1991), "Ellipsis and Higher-Order
Unification," Linguistics and Philosophy 14.4, 399–452.
Evans, Gareth (1980) "Pronouns," Linguistic Inquiry 11, 337–362.
Fiengo, Robert and Robert May (1994) Indices and Identity, MIT Press, Cambridge.
Frege, Gottlob (1892) "Über Sinn und Bedeutung," Zeitschrift für Philosophie und
philosophische Kritik 100, 25–50.
Grodzinsky, Yosef and Tanya Reinhart (1993) "The Innateness of Binding and
Coreference," Linguistic Inquiry 24, pp. 69–102.
Haas-Spohn, Ulrike (1995) Versteckte Indexikalität und subjektive Bedeutung, Berlin,
Akademie Verlag, Studia Grammatica 38.
Haïk, Isabelle (1984) "Indirect Binding," Linguistic Inquiry 15.2, 185–223.
Heim, Irene (1988) The Semantics of Definite and Indefinite Noun Phrases, New York:
Garland Press. [reprinted from 1982 Ph.D. thesis, Univ. of Massachusetts,
Amherst]
Heim, Irene (1990) "E-Type Pronouns and Donkey Anaphora," Linguistics and
Philosophy 13, 137–177.
244
Anaphora and Semantic Interpretation
Higginbotham, James (1980) "Anaphora and GB: Some Preliminary Remarks," in J.
Jensen ed. Cahiers Linguistiques d'Ottawa: Proceedings of NELS 10, Univ. of
Ottawa.
Higginbotham, James (1983) "Logical Form, Binding, and Nominals," Linguistic Inquiry
14, 395–420.
Higginbotham, James (1985) "On Semantics," Linguistic Inquiry 16, 547–593.
Higginbotham, James (1987) "On the Varieties of Cross-Reference," in GLOW X
Proceedings, Annali di Ca'Foscari, Venice, 123–142.
Higginbotham, James (1989) "Reference and Control," Rivista di Linguistica 1.2,
301–326.
Higginbotham, James (1994) "Anaphoric Reference and Common Reference," Linguistic
Inquiry.
Kaplan, David (1989) "Demonstratives," in J. Almog, J. Perry, H. Wettstein eds. Themes
from Kaplan, Oxford: Oxford Univ. Press. [circulated in manuscript form since
1977]
Kitagawa, Yoshihisa (1991) "Copying Identity," Natural Language and Linguistic
Theory 9.3, 497–536.
Lakoff, George (1972) "Linguistics and Natural Logic," in D. Davidson and G. Harman
eds. Semantics of Natural Language, Dordrecht: Reidel, 545–665.
Landman, Fred (1986) Towards a Theory of Information, Dordrecht: Foris.
Lasnik, Howard (1989) Essays on Anaphora, Dordrecht: Kluwer.
Lewis, David (1979) "Attitudes De Dicto and De Se," The Philosophical Review 88,
513–543.
Lewis, David (1984) "Individuation by Acquaintance and by Stipulation," in F. Landman
and F. Veltman eds. Varieties of Formal Semantics, Dordrecht: Foris, 219–243.
McKay, Thomas (1991) "Representing De Re Beliefs," Linguistics and Philosophy 14.6,
711–739.
Montague, Richard (1974) "The Proper Treatment of Quantification in Ordinary
English," in R. Thomason ed. Formal Philosophy, New York: Academic Press,
247–270.
Neale, Stephen (1990) Descriptions, Cambridge: M.I.T. Press.
Partee, Barbara and Emmon Bach (1984) "Quantification, Pronouns, and VP Anaphora,"
in J. Groenendijk, T. Janssen, M. Stokhof eds. Truth, Interpretation and
Information, Dordrecht: Foris, 99–130. [reprinted from Formal Methods in the
Study of Language, Univ. of Amsterdam, Mathematical Centre Tracts. 1981]
Pesetsky, David (1982) Paths and Categories, Ph.D. thesis, M.I.T., Cambridge.
Quine, W. V. O. (1956) "Quantifiers and Propositional Attitudes," The Journal of
Philosophy 53, 177–187.
Reinhart, Tanya (1983a) Anaphora and Semantic Interpretation, Chicago: The Univ. of
Chicago Press.
Reinhart, Tanya (1983b) "Coreference and Bound Anaphora: A Restatement of the
Anaphora Questions," Linguistics and Philosophy 6, 47–88.
Reinhart, Tanya (1991a) "Binding and Coreference Revisited," ms. Tel Aviv Univ.
Reinhart, Tanya (1991b) "Self-Representation," ms. Tel Aviv Univ. [presented at
anaphora workshop, Princeton Univ., Oct. 1990]
Rooth, Mats (1985) Association with Focus, Ph.D. thesis, Univ. of Massachusetts,
Amherst.
Sag, Ivan (1980) Deletion and Logical Form, New York: Garland Press. [reprinted from
1976 Ph.D. thesis, M.I.T., Cambridge]
Soames, Scott (1990) "Pronouns and Propositional Attitudes," The Proceedings of the
Aristotelian Society.
Stalnaker, Robert (1979) "Assertion," in P. Cole ed. Syntax and Semantics 9: Pragmatics,
315–332.
245
Irene Heim
Stechow, Arnim von (1982) "Structured Propositions," Univ. of Konstanz, report #59 of
the Sonderforschungsbereich 99.
Stowell, Timothy (1987) "Adjuncts, Arguments, and Crossover," ms. Univ. of California,
Los Angeles.
Irene Heim
Department of Linguistics and Philosophy
E39-245
MIT
Cambridge, MA 02139
U.S.A.
246